Skip to main content
HomeDocsDashboard
BUY DTAO
How RESI WorksDocumentationStatus Page
XLinkedInDiscord

How RESI works

RESI operates a 24-hour cycle competition on Bittensor Subnet 46.

The competition cycle​

Hour 0: Model submission​

Miners upload trained models to Hugging Face and register the hash on-chain. This timestamp proves ownership and submission time.

Hour 24.5: Data collection begins​

Validators scrape home sales from the previous 24 hours. This ensures no overlap between training data and validation data.

Hour 27: Validation begins​

Validators run inference on all submitted models using the new sale data. Models predict sale prices. Validators compare predictions to actual sale prices.

Performance measurement​

Validators measure model accuracy using R² (coefficient of determination) and potentially additional metrics. Top performers receive the highest weights.

Ongoing evaluation​

Previous winning models continue evaluation in each cycle. Yesterday's winner may have been submitted weeks ago. New models must outperform existing champions.

Participant roles​

Miners​

Goal: Create accurate property valuation models.

Process:

  1. Train models on historical property data
  2. Upload model files to Hugging Face
  3. Register model hash on Subnet 46
  4. Earn rewards when models outperform competitors

Requirements: Machine learning expertise, access to training data, computational resources for model development.

Validators​

Goal: Benchmark submitted models fairly and accurately.

Process:

  1. Register as validator on Subnet 46
  2. Pull daily home sales data
  3. Run inference on submitted models
  4. Calculate accuracy metrics
  5. Set weights based on performance
  6. Earn emissions for maintaining network integrity

Requirements: Computational resources for running inference, reliable uptime, technical capability to operate validator software.

Subnet operators​

Goal: Maintain validation dataset quality and timing.

Process:

  1. Ensure validation data is fresh and accurate
  2. Monitor validation cycle timing
  3. Maintain data scraping infrastructure

Application developers​

Goal: Access reliable property pricing data.

Process:

  1. Integrate RESI API into applications
  2. Request pricing predictions
  3. Receive predictions with model provenance and accuracy history

Stakers​

Goal: Earn returns by supporting validators.

Process:

  1. Stake TAO to validator hotkeys
  2. Earn proportional emissions based on validator performance

Anti-gaming measures​

Validators implement multiple checks:

  • Detect duplicate models from different miners (only first uploader rewarded)
  • Verify model execution in expected time bounds
  • Cross-check results between validators for consensus
  • Measure performance against established baselines

Model selection​

Validators use a threshold system. New models must achieve accuracy within a defined range of the current top model (typically within 10% of the previous day's best performer). This ensures efficient resource usage while discovering better models.

Future: Inspection integration​

RESI expands to include property inspections. Property owners upload images. AI models analyze images to assess condition. Inspection data adjusts appraisal prices to reflect actual property state.

This creates complete pricing: base appraisals from market data plus condition adjustments from visual inspection.

Next steps​